fix(reflect): fail on unusable tool calls instead of salvaging leaked text

[nicoloboschi]

Why

Reflect is driven entirely by structured tool calls. Some provider transports don't actually support function calling and silently strip the tool definitions from the request — most notably litellm's Vertex AI gpt-oss MaaS path, whose transformer drops tools/tool_choice when the model is flagged supports_function_calling=False. The model, never given the tools, answers in free text that mimics a done() payload, which lands in message.content with empty tool_calls.

The old reflect loop served that raw text as the answer. To cope, it grew an ever-expanding pile of regex/JSON "strippers" (_clean_done_answer, _unwrap_leaked_done_arguments, _strip_trailing_id_json_object, _clean_answer_text, _LEAKED_JSON_SUFFIX, _TRAILING_IDS_PATTERN, …) to claw the leaked memory_ids/observation_ids/directive_compliance siblings back out of the user-facing answer — each new provider quirk = another pattern (#2222, #2345, #2757, and the pending #2972).

Verification

Reproduced against real gpt-oss:120b-cloud:

• Over Ollama, the done tool call parses cleanly every time — structured tool_calls, empty content, no leak.
• The leak is transport-specific: litellm's vertex_ai/vertex_ai_partner_models/gpt_oss/transformation.py strips the tool params for gpt-oss MaaS. It's not a model limitation — the same model tool-calls perfectly on a transport that supports it.

What changed

Fail loudly instead of salvaging untooled text:

• Track whether the model ever produced a tool call reflect could parse. If it never does (the stripped-tools case), raise ReflectToolCallError → HTTP 500 (the request is valid; the server's configured model can't do the job) with a clear message (provider, model, response snippet).
• Keep the done tool. _process_done_tool now trusts args["answer"] verbatim — a parsed tool call can't bleed its sibling id fields into the answer string.
• A model that did tool-call and later stops with text is a legitimate completion and still routes through the clean forced-final synthesis path (no leak — it's a fresh no-tools call).
• Delete the entire strip zoo, including _clean_answer_text / _DONE_CALL_PATTERN. The forced-final paths return the model's prose directly — tools are disabled there, so there's no tool syntax to strip.

No static supports_function_calling gate — reflect just tries and fails. It's reliable because iteration 0 forces a named tool choice, so any compliant provider yields a tool call and never trips the guard; only a transport that strips tools fails, fast, on turn 0.

Note for downstream clients

Reflect can now return HTTP 500 for a model/transport that can't drive tool calls, where it previously returned 200 with leaked or hollow text. No route/request/response schema changed, so no OpenAPI/client regeneration is required.

Checks

• ruff check / ruff format --check / ty check — clean
• pytest tests/test_reflect_agent.py (non-LLM) — 39 passed
  • added test_no_tool_call_ever_raises_tool_call_error and test_stop_after_evidence_uses_forced_final_synthesis
  • removed tests for the deleted salvage helpers
• test_directive_not_echoed_on_empty_bank errors locally on a shared-pg0 port conflict (pre-existing, environment), unrelated to this change

Supersedes the answer-salvage approach in #2972.